A Pedestrian Target Detection Network Based on Attention Mechanism and Multi-scale Feature Fusion

An improved YOLOv7 pedestrian target detection algorithm is proposed to solve the problems of low target detection accuracy caused by large number of pedestrians, occlusion, scale change and light shade in pedestrian target detection. Based on YOLOv7, Triplet Attention (TA) is integrated into ELAN (Efficient layer aggregation network) structure. For the head part, the fusion thought of Gold-YOLO is used to improve the feature fusion stage of YOLOv7, and the ELAN structure of the head part is improved by incorporating the idea of Triplet attention mechanism and residuals. The proposed model is known as GDA-YOLO. Due to the single pedestrian characteristics of a single dataset, in order to increase the diversity of data in the dataset, multiple public datasets form a mixed dataset, and verify the performance of the GDA-YOLO model on this mixed dataset. The experimental results show that the improved detection algorithm achieves 89.25% mAP@0.5 and 95.6 GFLOPs, and the improved model achieves higher detection accuracy and better detection results compared to the current popular detection models.